Apprenticeship
The Academy
Seven levels, the way a watchmaker learns: first follow instructions, then plan, then build your own tools — and finally sit at the real bench.
Before level 1: add the connector
Robot MCP connector address:
Sign-in required. Tower of Hanoi and free-bench simulation controls are enabled. Training and the physical arm are not connected yet.
Free-bench track
Teach the robot
You own intent and success criteria. Claude owns the how. The app measures.
1. Baseline
Write an intent and measurable checklist in Lab. Start without a skill or guardrails.
Use Robot MCP. List my mission cards and ask which to use. Read that card. Start free-bench with its card_id, label baseline and open get_camera_frame ONCE. Execute the intent without a saved skill or guardrails. End the run and read get_run_report. Explain each failed criterion.
Good looks like: An honest scorecard; failures are evidence, not something to hide.
2. Skill
Ask Claude to turn what it learned into a repeatable plan, with observation and verification.
Read my mission card and baseline report. Write a reusable Claude skill for the intent: plan, observe, act and verify. Show the skill text so I can save it in Claude. Run it on a fresh free-bench session labelled skill v1, open the camera once, end and compare_runs with baseline.
Good looks like: A concrete reusable skill and a better measured run, not just a confident claim.
3. Guardrails
Add server-enforced protection where the skill can still break your criteria.
Read the card and compare_runs. Propose guardrails using set_guardrails with unique ids, reasons and block or warn modes. Show why each protects my criteria. After my approval set them; start a new run labelled skill v1 + guardrails, open the camera once, execute the skill, end and compare. Report blocks and warnings honestly.
Good looks like: Each guardrail has a purpose; blocked attempts remain visible in the scorecard.
4. Iterate
Tighten action or time limits, then refine until every criterion passes twice consecutively.
Read my updated mission card and compare_runs. Refine the skill for the tighter criteria. Run fresh sessions and compare their reports. Stop when every measurable criterion passes twice in a row; flag written criteria for my review. Keep one camera player per session. Show the final skill and the evidence.
Good looks like: Two consecutive passing scorecards, with manual review for written criteria.
- 1
Command
FreeStart by giving the robot direct, single-step instructions. Learn the tools: look, move_to, pick and place, and how Claude reports what it sees.
Use the Robot MCP connector. Look at the bench, describe every object and its grid cell, then pick up the red cube and put it on C3.
- 2
Plan
FreeGive a goal, not steps. Claude must break it into a sequence, check the scene between moves and recover when something slips.
Build a tower on E1 with all the cubes, largest at the bottom. Plan first, show me the plan, then execute and verify after each placement.
- 3
Program a skill
FreeTurn a reliable sequence into a named, reusable skill with parameters. Skills are saved to your account and can be called in later sessions.
Write a skill called stack(object, cell) that hovers, descends, grips, lifts and places with a check. Test it on two cubes and save it.
- 4
Teach by demonstration
FreeRecord a demonstration from the scripted controller and ask Claude to generalise it into a skill that works from new starting positions.
Watch the recorded demo 'pour beads into cup'. Describe the key phases, then reproduce it with the cup moved to B5.
- 5
Train and test
ProLaunch a training job on randomised worlds, then evaluate the learned policy on held-out seeds. Compare it against your hand-written skill.
Start a training job for 'sort cubes by colour' on 50 random worlds, then test on seeds 900–909 and report the success rate.
- 6
Coach
ProSupervise the robot while a policy runs. Interrupt, give corrective advice and let Claude decide when to fall back on a skill.
Run the sorting policy. If any cube is misplaced, stop, explain what went wrong and fix it with the stack skill.
- 7
Graduate to the real arm
ProBring your best skill to the real SO-101 in Geneva. Expect friction, lighting and slight calibration drift — the true test of robustness.
Book a real-arm session. Run my stack skill on the physical bench, film it, and summarise any differences from the simulation.